git.ipfire.org Git - thirdparty/kernel/stable-queue.git/blob

   1 From foo@baz Tue Apr 24 15:29:20 CEST 2018
   2 From: Vitaly Kuznetsov <vkuznets@redhat.com>
   3 Date: Thu, 25 Jan 2018 16:37:07 +0100
   4 Subject: x86/kvm/vmx: do not use vm-exit instruction length for fast MMIO when running nested
   5
   6 From: Vitaly Kuznetsov <vkuznets@redhat.com>
   7
   8
   9 [ Upstream commit d391f1207067268261add0485f0f34503539c5b0 ]
  10
  11 I was investigating an issue with seabios >= 1.10 which stopped working
  12 for nested KVM on Hyper-V. The problem appears to be in
  13 handle_ept_violation() function: when we do fast mmio we need to skip
  14 the instruction so we do kvm_skip_emulated_instruction(). This, however,
  15 depends on VM_EXIT_INSTRUCTION_LEN field being set correctly in VMCS.
  16 However, this is not the case.
  17
  18 Intel's manual doesn't mandate VM_EXIT_INSTRUCTION_LEN to be set when
  19 EPT MISCONFIG occurs. While on real hardware it was observed to be set,
  20 some hypervisors follow the spec and don't set it; we end up advancing
  21 IP with some random value.
  22
  23 I checked with Microsoft and they confirmed they don't fill
  24 VM_EXIT_INSTRUCTION_LEN on EPT MISCONFIG.
  25
  26 Fix the issue by doing instruction skip through emulator when running
  27 nested.
  28
  29 Fixes: 68c3b4d1676d870f0453c31d5a52e7e65c7448ae
  30 Suggested-by: Radim Krčmář <rkrcmar@redhat.com>
  31 Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
  32 Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
  33 Acked-by: Michael S. Tsirkin <mst@redhat.com>
  34 Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
  35 Signed-off-by: Sasha Levin <alexander.levin@microsoft.com>
  36 Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
  37 ---
  38  arch/x86/kvm/vmx.c |   16 +++++++++++++++-
  39  arch/x86/kvm/x86.c |    3 ++-
  40  2 files changed, 17 insertions(+), 2 deletions(-)
  41
  42 --- a/arch/x86/kvm/vmx.c
  43 +++ b/arch/x86/kvm/vmx.c
  44 @@ -6765,7 +6765,21 @@ static int handle_ept_misconfig(struct k
  45         if (!is_guest_mode(vcpu) &&
  46             !kvm_io_bus_write(vcpu, KVM_FAST_MMIO_BUS, gpa, 0, NULL)) {
  47                 trace_kvm_fast_mmio(gpa);
  48 -               return kvm_skip_emulated_instruction(vcpu);
  49 +               /*
  50 +                * Doing kvm_skip_emulated_instruction() depends on undefined
  51 +                * behavior: Intel's manual doesn't mandate
  52 +                * VM_EXIT_INSTRUCTION_LEN to be set in VMCS when EPT MISCONFIG
  53 +                * occurs and while on real hardware it was observed to be set,
  54 +                * other hypervisors (namely Hyper-V) don't set it, we end up
  55 +                * advancing IP with some random value. Disable fast mmio when
  56 +                * running nested and keep it for real hardware in hope that
  57 +                * VM_EXIT_INSTRUCTION_LEN will always be set correctly.
  58 +                */
  59 +               if (!static_cpu_has(X86_FEATURE_HYPERVISOR))
  60 +                       return kvm_skip_emulated_instruction(vcpu);
  61 +               else
  62 +                       return x86_emulate_instruction(vcpu, gpa, EMULTYPE_SKIP,
  63 +                                                      NULL, 0) == EMULATE_DONE;
  64         }
  65
  66         ret = kvm_mmu_page_fault(vcpu, gpa, PFERR_RSVD_MASK, NULL, 0);
  67 --- a/arch/x86/kvm/x86.c
  68 +++ b/arch/x86/kvm/x86.c
  69 @@ -5699,7 +5699,8 @@ int x86_emulate_instruction(struct kvm_v
  70                  * handle watchpoints yet, those would be handled in
  71                  * the emulate_ops.
  72                  */
  73 -               if (kvm_vcpu_check_breakpoint(vcpu, &r))
  74 +               if (!(emulation_type & EMULTYPE_SKIP) &&
  75 +                   kvm_vcpu_check_breakpoint(vcpu, &r))
  76                         return r;
  77
  78                 ctxt->interruptibility = 0;